ChatGPT vs Claude plans for companies: which subscription to buy, and for whom (2026 buyer's guide)

OpenAI and Anthropic each sell seven plans, and the names hide the decision. For a company, the choice comes down to four shapes: a personal plan (ChatGPT Free, Go, Plus, Pro; Claude Free, Pro, Max) that trains on chats by default and has no admin, a standard workspace seat at about £16 to £20 a month (ChatGPT Business, Claude Team), a premium seat at about £80 to £100 for the people who use the flagship model all day, and an Enterprise contract for single sign-on, audit logs and, on OpenAI's side, European data residency. Most teams we meet should buy standard seats, count who hits the ceiling for a month, then upgrade those people. The full comparison table and the use case for each plan are below.

The spreadsheet on the COO's desk

You've got a column of expense claims. ChatGPT Plus, £19.20, eleven times. Claude Pro, £18, four times. One Max at £90 for the head of product, who says she needs it. A ChatGPT Pro at £192 nobody can explain. The CFO wants one line instead of seventeen, IT wants to know where the client data went, and you've been asked to pick a plan by the end of the month.

The vendor pricing pages don't help much. OpenAI lists Free, Go, Plus, Pro, Business, Enterprise and Edu. Anthropic lists Free, Pro, Max, Team, Enterprise, plus two seat types inside Team. Fourteen options, two of which are the same price with different things inside.

We help companies choose and roll out these tools for a living, and most of the ones who call us have already bought the wrong tier. Usually too much, sometimes too little, almost always without counting. This guide covers every plan from both providers as of 12 September 2026, the need each one solves, and the rule we use to pick.

Every ChatGPT and Claude plan, side by side

The two vendors bill UK companies differently, and the table reflects what you'll see at checkout. OpenAI prices its personal plans in pounds before VAT and adds 20% at checkout (OpenAI, multi-currency billing); its Business seats are still priced in dollars, converted by your card and then taxed, so the pound figures for those rows are approximate at a rate of about 0.80. Anthropic shows UK visitors sterling prices with VAT already included (AIToolsReview's September 2026 UK guide); its Enterprise plan stays in dollars. A VAT-registered company reclaims the 20% either way, as long as the invoice is in the company's name. Where a vendor doesn't publish a price, the table says so and cites what customers report.

Two rows in that table decide most purchases. The "trains on your data" column splits every personal plan from every workspace plan, on both sides. And the flagship column shows the one real difference between a ChatGPT Business seat and a Claude Team seat at nearly the same money: OpenAI puts a slice of its best model inside the standard seat, Anthropic sends you a bill for it.

OpenAI's plans, one by one

ChatGPT Free (£0)

Since August 2026, Free users get unlimited text chats on GPT-5.6 Luna, the lightest tier of the GPT-5.6 family, with images, file uploads and voice capped separately. OpenAI started showing ads to Free users in the US, with other countries following (The AI Career Lab's plan comparison). The Work agent and Codex run on GPT-5.6 Terra with limited desktop access.

The need it solves: a colleague who wants to see what the fuss is about before anyone spends money. At work, that's where it should end. Training is on by default, there's no admin, and an ad next to a client brief is a conversation you don't want to have with the client.

ChatGPT Go (about £6 a month)

Go keeps the Luna model and the ads, raises the limits, and adds a lighter form of deep research. No Sites, no Sol, no GPT-6. OpenAI built it for consumers in price-sensitive markets, and £6 doesn't buy the ads off.

The need it solves: a personal account for someone who chats a lot and codes never. We haven't found a company use for it. The £10 gap to Plus is where the work features start.

ChatGPT Plus (£16 a month plus VAT, £19.20 at checkout)

Plus is the first tier that feels like the product people talk about. GPT-5.6 Sol as the default model, ad-free, deep research, agent mode, the ChatGPT Work agent, Codex, Sites for small web apps and dashboards, custom GPTs, Projects, expanded memory and image generation with Images 2.5. GPT-6 Astra is on Plus, but only inside ChatGPT Work and Codex on a limited allowance, and not in regular chat (Notebookcheck).

The need it solves: one professional, working alone, on their own material. A freelancer, a founder before the first hire, a consultant who never pastes client data. For a company it's the plan that appears on expense claims, and that's the problem covered below.

ChatGPT Pro (£80 or £160 a month plus VAT)

Pro is Plus with the meter opened up: five times the usage at £80 (£96 with VAT), twenty times at £160 (£192), plus the Pro-level reasoning modes. Since 3 September 2026, that includes GPT-6 Pro in regular chat, which is GPT-6 Astra with more thinking time: 50 messages a week on the £80 plan, 200 on the £160 plan (The Decoder).

The need it solves: a single person who runs agents for hours, a developer living in Codex, an analyst who hits the Plus ceiling by lunch. It's still a personal account, with training on by default and no admin or workspace behind it. If that person works in a company, a premium Business seat costs about the same £80 a month on annual billing and comes with the admin console. The £160 tier is for someone whose usage a company would rather see on an invoice than on an expense claim.

ChatGPT Business, standard seat ($25 a month, or $20 annual: about £20 or £16 plus VAT)

Business is the workspace plan, bought by card without a sales call. OpenAI still prices it in dollars for UK customers, so the pound figure moves with the exchange rate and VAT is added on top (AI Build Group, OpenAI's UK channel partner). Workspace content and connector data are excluded from training by default, with nothing to opt out of (OpenAI Help Centre). You get an admin console, SSO, shared Projects and custom GPTs, connectors to Drive, SharePoint, Slack and the rest, the Work agent, Codex, deep research, image generation, and the GPT-6 family for daily use. GPT-6 Pro is off until an admin enables it, then capped at 15 messages a month per standard seat, and usage sits inside a five-hour window.

The need it solves: any team of two or more that touches company or client material. A sales team drafting proposals from CRM exports. A marketing team running Images 2.5 on real campaigns. A finance team that wants the same tool for everyone and one invoice. This is the default we recommend on the OpenAI side, and the fifteen GPT-6 Pro messages are enough to find out who needs more.

ChatGPT Business, premium seat ($125 a month, or $100 annual: about £100 or £80 plus VAT)

OpenAI added premium seats to Business in August 2026: five times the usage of a standard seat, no five-hour window, and 50 GPT-6 Pro messages a week (OpenAI's announcement). Standard and premium seats mix in the same workspace, so you pick per person.

The need it solves: the three people in a 30-person workspace who run the Work agent all afternoon and hit the five-hour cap at 3pm. Developers on Codex, an ops lead rebuilding reporting, a strategy analyst on deep research. Those three get premium and the other 27 stay standard, because a premium seat charges about £100 a month whether the person reaches the cap or not.

ChatGPT Enterprise (quote)

OpenAI doesn't publish an Enterprise price. Buyers in 2026 report contracts between roughly $45 and $75 a seat a month, annual, which is about £36 to £60 before VAT, with the lower end for larger or multi-year deals (Layer3 Labs' Enterprise pricing guide; treat these as reported, not list). The seat carries higher limits than Business, with extra usage billed through a credit rate card where a GPT-6 Pro message costs 50 credits (OpenAI rate card). What you're paying for is the control layer: SAML SSO, SCIM provisioning, role-based access, audit logs, a compliance API, enterprise key management, a HIPAA agreement where needed, a customer success manager, and data and inference residency in Europe, the UK, the US and seven other regions at no extra cost (OpenAI, data residency).

The need it solves: a company whose security team has a checklist, and a French or German legal department that wants the data to stay in the EU. Below a few hundred seats, Business does the same work for less, and it already has SSO, which is the feature most mid-sized buyers think they're paying Enterprise for.

ChatGPT Edu (quote)

Edu is the Enterprise feature set priced for universities, with the same residency and no-training terms. If you're a company, it's not for you. If you run a business school, it's the one to ask about.

Anthropic's plans, one by one

Claude Free (£0)

Free gives you the chat product with tight session caps, no Fable models, and the same training question as every consumer plan. Since October 2025, Free, Pro and Max users choose whether their chats train future models. The setting defaults to on, and anyone who leaves it there accepts five-year retention of that data (Anthropic, consumer terms update).

The need it solves: a first look. Same verdict as ChatGPT Free: fine at home, wrong for a client file.

Claude Pro (£18 a month, or £15 annual, VAT included)

Pro is the individual plan and it carries more than its ChatGPT counterpart for about the same money (Anthropic's UK page shows £18 a month with VAT included, or £180 a year): Opus 5 and Sonnet 5 inside the plan, unlimited Projects, connectors through MCP, the Google Workspace integration, Claude Code in the terminal, and Claude Cowork on the desktop for agentic file and task work. Usage is metered on a five-hour session and a weekly cap shared across chat, Code and Cowork. Fable 5.1, the flagship, runs on usage credits billed on top of the subscription.

The need it solves: one person who writes long, reads long documents, or codes. A lawyer reviewing 200-page contracts alone, a solo developer, a founder drafting the board pack. It's still a personal account, with training on by default and nobody able to check it. The usage limits guide explains why Pro users hit the ceiling and what to do about it.

Claude Max (£90 for 5x, £180 for 20x, VAT included)

Max multiplies Pro's limits by five or twenty and includes Fable 5.1 inside the plan, up to half the weekly limit, with the rest on Opus 5 and Sonnet 5 (Anthropic, Fable models on your plan). Priority access at busy times and early access to new features come with it.

The need it solves: a developer running Claude Code all day, a consultant who lives in Cowork, anyone whose Fable 5.1 usage credits on Pro started costing more than the £72 gap. The same caveat as ChatGPT Pro applies: it's a personal account. In a company, a premium Team seat costs about the same £90 a month on annual billing and gives the admin a console.

Claude Team, standard seat (about £22 a month, or £18 annual, VAT included)

Team is the workspace plan, from two seats since 20 July 2026 and up to 150, bought by card. Anthropic's UK page lists the standard seat at £18 a month on annual billing with VAT included, which is £15 net for a VAT-registered company; the monthly figure is about £22. Workspace data is excluded from training as a contractual term, with 30-day default retention. A standard seat carries 1.25 times Pro usage, Opus 5 and Sonnet 5, Claude Code on every seat, shared Projects, Skills, connectors, and an admin console with SSO (Anthropic, Team plan). Fable 5.1 is pay-as-you-go: usage credits billed on top of the seat, per person, as they use it.

The need it solves: the same teams as ChatGPT Business, with a tilt towards document-heavy and code-heavy work. Legal, finance, product, engineering, agencies writing proposals. We covered the UK setup in Claude.ai 101 for business teams. The pay-as-you-go flagship is a feature for a team where three people need Fable 5.1 and 20 don't: the light users cost nothing extra.

Claude Team, premium seat (about £112 a month, or £90 annual, VAT included)

Premium raises the multiplier to 6.25 times Pro, puts Fable 5.1 inside the limits under the same half-of-weekly rule as Max, and adds Claude Cowork for agentic desktop work. Seats mix in one workspace.

The need it solves: the people whose monthly usage credit bill on a standard seat crossed £80, and anyone the team wants on Cowork. Engineering leads, a product manager building internal tools, an analyst running research agents. Same rule as OpenAI: buy it for named people, after a month of counting.

Claude Enterprise ($20 a seat, annual, plus usage: about £16 plus VAT)

Enterprise moved to a usage-based shape during 2026. The seat is $20 a month billed annually, about £16 before VAT, and covers the platform. Enterprise is one of the two Claude plans still priced in dollars for UK buyers. Every token your people spend, in chat, Claude Code or Cowork, is billed on top at standard API rates, with no allowance in the seat (Justin McKelvey's Enterprise pricing breakdown). Self-serve starts at 20 seats; through sales, 50. The seat buys SSO, SCIM, audit logs, custom retention, a compliance API and an analytics API, customer-managed encryption keys, a US-only inference option, and HIPAA-ready configuration for eligible companies.

Two things a European buyer should know. Claude Enterprise doesn't offer EU data residency inside the chat product; storage is US-hosted and GDPR is handled through standard contractual clauses and a data processing agreement, with regional endpoints available through AWS Bedrock, Google Cloud and Microsoft Foundry for API traffic (Compound's EU hosting guide). And Anthropic's Covered Models policy keeps prompts and outputs for its most capable models for 30 days on every platform, Enterprise included, which we unpacked in the zero data retention comparison.

The need it solves: a company with a security checklist, 20 or more seats, and a finance team willing to model usage. The metered shape is the honest one for a heavy-user population: you pay for what runs. It's a bad shape for a company that can't yet say who its heavy users are, because the first invoice will tell them.

Personal plans at work: the expense-claim problem

The seventeen subscriptions on that spreadsheet share one property. Each one is a consumer account, and each consumer account, on both providers, trains on conversations unless the individual found the toggle. On ChatGPT it lives under Settings, Data Controls. On Claude it's the choice presented at signup and again under Privacy. Nobody at the company can see who flipped it.

The other half of the problem is the admin console that doesn't exist. No list of who's using what. No way to remove access when someone leaves. No connector policy. No data processing agreement, because the consumer terms aren't one.

The arithmetic ends the debate in most rooms. Eleven Plus accounts on expenses cost £211 a month with VAT, on invoices in eleven employees' names. Eleven ChatGPT Business seats on annual billing cost about £176 a month before VAT, on one invoice in the company's name, which your accountant will prefer for the reclaim. Same money or less, and the second one comes with the training question answered in the contract and a console that shows the ops lead who logged in this week. The Claude version is the same: Pro at £18 against a Team seat at £18 on annual billing.

"Most of the shadow AI I find in UK companies isn't people using banned tools. It's twelve people expensing a personal Pro subscription and pasting client work into an account with the training toggle still on. The fix costs less than the risk." Toni Dos Santos, in Claude.ai 101 for business teams.

The one honest exception is the Max 20x or Pro £160 user. A company can't buy the top personal tier as a seat; the premium seat tops out at the equivalent of the £80 or £90 tier. If one person genuinely needs 20x usage, they'll either stay personal with the toggle off and a signed acknowledgement, or move to Enterprise where usage is metered. We've seen both. The first is cheaper and the second is cleaner.

Which plan for which need

The table below is the one we draw on the whiteboard. Start from the situation, read across.

ChatGPT or Claude: how we'd choose between the providers

The plan shapes are now nearly identical, so the provider decision comes down to what's inside the seat and what your people do all day. Our position, which you can argue with:

Broad teams go to ChatGPT Business. Sales, marketing, HR, ops, customer support, people who need images, voice, a familiar interface and the flagship model in the standard seat without a usage bill. OpenAI's bundle in a standard seat is wider: image generation, agent mode, deep research, Codex, custom GPTs, connectors, plus 15 GPT-6 Pro messages a month. Claude has no image generation at all, and that alone settles it for most marketing teams.

Deep work goes to Claude Team. Long documents, code, agentic desktop work through Cowork and Skills, teams that write for a living. In our experience, the teams that pick Claude do it for the data room and the codebase, then stay for Projects and Skills. Pay-as-you-go Fable 5.1 on a standard seat is a feature here: the two people who need the flagship pay for it, the rest don't.

Residency on the checklist means ChatGPT Enterprise. If your DPO's checklist says EU data at rest and EU inference in the chat product, Anthropic can't tick that box today outside the cloud APIs. That's the single clearest technical difference at the Enterprise tier.

We wouldn't buy either on the model benchmark. The two flagships trade places every quarter. When we checked on 6 September, Artificial Analysis had Fable 5.1 at 57 and GPT-6 Astra at 55 on its Intelligence Index, a gap no team feels on a budget reconciliation. We wrote up why the annual pick-a-model exercise fails in AI model fatigue in the enterprise, and why a migration rarely starts from a task in the Claude to ChatGPT switching guide.

Both at once is a real option, and the two-standard-seat version costs about £35 a person a month on annual billing. It works when marketing is on ChatGPT and engineering is on Claude, each with its own admin. It fails when everyone has both and nobody owns either. AI spend visibility covers how to see that from the admin consoles.

What we do before a company signs

The rule is short: standard seats first, count for 30 days, then buy premium for named people. This is how it ran the last few times, with the specifics removed.

The ops lead exports the expense claims and lists every personal subscription with the owner's name. That list becomes the first seat list, and the personal accounts get cancelled on the day the workspace opens, with a two-line note about the training toggle. Nobody argues, because the seat costs the same.

Everyone starts on a standard seat. On ChatGPT, the admin enables GPT-6 Pro so the 15 messages a month are visible. On Claude, usage credits for Fable 5.1 are switched on with a monthly cap per person. Then the team runs its real workflows for a month: the proposal, the monthly close, the sprint, the campaign. Cycle time on each gets written down on day one, which is the baseline the budget conversation needs later.

At the end of the month the admin console answers the question the vendor pricing page couldn't. Who hit the five-hour window on ChatGPT. Whose Fable credits on Claude crossed £80. Those people get premium seats. In our experience they're a small minority of the workspace, and you should count yours before trusting anyone's ratio, ours included. A 40-person ChatGPT Business workspace with four premium seats runs at about £900 a month before VAT on annual billing. The same workspace bought as 40 premium seats, which is what the salesperson suggested, runs at about £3,200.

Then the second question, the one we actually get paid for: the seat is live, so what changed? A team on the right plan with no workflow change looks exactly like the team on expense claims, at the same price. That's the part covered in our procurement guide and the reason every engagement we run ends with a measured change in cycle time on the workflows that were baselined on day one.